Anthropic enhances Claude: Fable 5.1 takes the benchmark lead

Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1. According to Anthropic, the generally available Fable 5.1 has improved, particularly in long-running agent tasks, coding, and knowledge work.

In Anthropic’s own tests, the leap is particularly noticeable in agentic tasks: on Terminal-Bench-Science, a test for complex scientific workflows, Fable 5.1 increases from 24.7 to 52.6 percent. On Terminal-Bench 4.0, which tests agentic coding tasks in a terminal environment, the model improves from 42.0 to 55.8 percent. On AutomationBench, which simulates multi-step office tasks across various applications, Fable 5.1 increases from 17.1 to 31.4 percent. Anthropic also reports further, mostly smaller, gains in knowledge work, computer control, and general reasoning.

External tests by artificial analysis support the performance leap. With maximum reasoning power, Fable 5.1 achieves a new top score of 66 points in the Intelligence Index, leading the ranking. For comparison: Fable 5 scores 62 points, Opus 5 reaches 63, and GPT-5.6 Sol scores 61 points. Fable 5.1 also achieves the highest values to date in Terminal-Bench v2.1 and SciCode. However, in other tests for agentic knowledge work, it is practically on par with Opus 5.

Cheaper cache access with still high costs

Fable 5.1 remains one of the most expensive top models. Regular API prices remain unchanged compared to Fable 5, at $10 per million input tokens and $50 per million output tokens. This makes Fable 5.1 twice as expensive as Opus 5 for normal input and output tokens. Anthropic itself recommends Opus 5 for most applications and Fable 5.1 primarily for particularly demanding and lengthy tasks.

However, repeated use of already processed content becomes significantly cheaper. Cache Reads now cost only $0.25 instead of $1 per million tokens, making them half as expensive as with Opus 5 for Fable 5.1. This particularly benefits workloads where already processed context is reused. Anthropic estimates that total costs will decrease by around 25 percent for typical workloads and up to about 45 percent for highly agentic workloads compared to Fable 5.

However, the tests by artificial analysis show that higher token consumption can offset this saving. With maximum reasoning power, Fable 5.1 generated about 1.7 times more output tokens than Fable 5. Consequently, an intelligence index task cost an average of $3.76. This is 20 percent more than with Fable 5 and about 60 percent more than with Opus 5.

Mythos 5.1 with less strict safety mechanisms

Anthropic has also revised its safety mechanisms. Fable 5.1 can now identify vulnerabilities in source code, while riskier cyber tasks such as penetration testing, exploit generation, and binary vulnerability scanning remain restricted. Due to the more precise safety mechanisms, the cyber safeguards in Claude Code intervene, according to Anthropic, on average about 60 percent less often than with Fable 5.

Mythos 5.1, accessible only to verified users and organizations, has less strict cyber safeguards than Fable 5.1. In internal security tests, it showed the strongest cyber capabilities of all Anthropic models released to date, but the company assesses it to be still below the threshold for the higher cyber risk category.

Fable 5.1 and Mythos 5.1 are the first new Claude models with Anthropic’s text watermarking. An initially limited-access detection API is intended to enable media, fact-checkers, and researchers, among others, to identify such texts.

(dahe)